| Feature pooling networks have recently become an essential core of many state-of-the-art visual recognition systems. Pooled image features, the output of these networks, allow a classifier to better relate limited amount of training examples of an object to unseen test cases in the face of highly uncertain viewing condition involved. However, despite their impressive performance, current feature pooling networks are designed only with simple translational invariance. In this thesis, we set out to increase the potency of these feature pooling networks by enhancing their view invariance.;We pursue the synthesis of complex invariances in object features through a direct modelling approach, an approach so far not pursued in the research community to the best of our knowledge. We begin with view-invariant K-means (VIK) algorithm, a novel dictionary learning algorithm that incorporates specific invariances directly into object features. With limited samples on the CIFAR-10 dataset, the algorithm led to an accuracy of 70.5%, competitive to the then best published results. A randomized version of the design, mix-VIK, manages to combine multiple complex invariances while keeping basis diversity and achieves an accuracy of 63.7% on the STL-10 dataset and 72.6% on CIFAR-10 dataset with limited samples, a state-of-the-art level. Finally, we present a receptive field transform method that challenges traditional receptive field design, which gives further accuracy gains on STL-10 and CIFAR-10, with accuracies of 64.9% and 74.3% respectively.;Despite its simplicity, our approach repeatedly demonstrated superior performance over competing approaches in complex invariances. We believe our studies here provide a good framework for modeling view invariance for high-level object features in practice, and there are important insights to be drawn from our study on the design of next generation feature pooling networks. |
